What Influences Readers' and Writers' Perceived Necessity of AI Disclosure?

This study investigates the factors influencing the perceived necessity of AI disclosure in writing through a vignette study of 727 participants, revealing that readers demand more transparency than writers and that judgments are significantly shaped by the AI's replaceability, directness, and the writer's intentionality, while writing effort has no significant impact.

The Gist

The Big Question: Who Needs to Know About the AI?

Imagine you are baking a cake. Sometimes you bake it entirely from scratch. Other times, you use a pre-made mix, or maybe you ask a friend to help you frost it.

Now, imagine a new rule is being debated: "Do you have to tell people if you used help (or a machine) to make your cake?"

This paper asks a very specific question: When do people (the readers) and the bakers (the writers) think it is necessary to admit that AI was used?

The researchers found that readers and writers often see this situation through very different lenses, and the answer depends heavily on how the AI was used.

---

The Three Main Ingredients of the Study

To figure this out, the researchers set up a "tasting menu" of 727 different scenarios (called vignettes). They mixed and matched three main ingredients to see how they changed people's minds:

1. The Perspective (Who is talking?): Are you the Reader (eating the cake) or the Writer (baking it)?
2. The Purpose (What is the cake for?): Is it for a serious job interview (Evaluation), a school lesson (Learning), or just a fun story for a blog (Entertainment)?
3. The Procedure (How was the cake made?): This is the most complex part. They looked at four specific ways AI was involved:
   • Replaceability: Could the writer have made this cake without the AI? (If the answer is "No," the AI was essential).
   • Effort: Did the writer still work hard, or did the AI do almost everything?
   • Intentionality: Did the writer give the AI a strict recipe and steer the ship, or did they just say "Make me a cake!" and let the AI take the wheel?
   • Directness: Did the writer copy the AI's output directly onto the plate, or did they use it just as a reference?

---

The Surprising Results

Here is what the study found, translated into plain English:

1. The "Self-Serving" Gap

The Finding: Readers are much more likely to say, "Yes, you must tell us!" than writers are.
The Analogy: Imagine a magician. The audience (readers) wants to know if a trick is real or if it's a machine doing the work. But the magician (writer) might think, "I still performed the show; the machine was just a prop." The study suggests writers often feel less need to disclose because they still feel like the "owner" of the work, even if they used AI.

2. The "Purpose" Myth

The Finding: Surprisingly, it didn't matter much why the text was written.
The Analogy: You might think people care more about honesty when the text is for a serious job application (Evaluation) than for a funny blog post (Entertainment). But the study found that people's need to know about AI usage was roughly the same whether it was for school, work, or fun. The "stakes" didn't change the need for disclosure.

3. The "How" Matters Most (Procedural Factors)

The study found that how the AI was used changed everything. Disclosure was seen as more necessary when:

• The AI was Irreplaceable: If the writer couldn't have done it alone without the AI, people felt it was crucial to admit it.
• The AI Took the Wheel (Low Intentionality): If the writer just typed "Write a story" and let the AI decide the plot, readers felt a strong need to know.
• The AI Output was Used Directly: If the writer just pasted the AI's text with very few changes, disclosure was seen as necessary.

The Twist on Effort:
The researchers thought that if a writer put in very little effort (letting the AI do 90% of the work), people would demand disclosure. They were wrong. The amount of effort the writer put in didn't significantly change whether people thought disclosure was necessary. It wasn't about how tired the writer was; it was about how much the AI actually did.

4. The "Steering" Paradox

The Finding: This is the most confusing part.

• Readers think: "If you let the AI drive the car (Low Intentionality), you must tell us!"
• Writers think: "If I let the AI drive the car, I feel less need to tell anyone."
• The Analogy: It's like a driver who says, "I didn't really drive; the GPS drove." The driver feels less responsible, but the passengers (readers) feel like they need to know the car was on autopilot.

---

What Does This Mean for the Future?

The paper suggests that because writers and readers see things differently, we can't just rely on writers to "do the right thing" on their own.

The Solution: We need better tools.
Instead of just asking writers to "disclose AI use," platforms should help them explain how they used it.

• Don't just say: "I used AI."
• Do say: "I used AI to generate the outline, but I wrote the sentences myself," or "I let the AI write the whole draft, and I just edited it."

The paper suggests designing tools that automatically draft these explanations for writers, making it easier for them to be transparent in a way that matches what readers actually care about (like whether the AI did the heavy lifting or just helped with a few words).

Summary

• Readers want to know more about AI use than Writers do.
• Why the text was written (school vs. fun) doesn't change the need for disclosure.
• How the AI was used matters most: If the AI did the heavy lifting, took the lead, or was copied directly, people feel disclosure is necessary.
• Effort doesn't matter as much as we thought.
• Writers often feel less need to disclose when they let the AI take the lead, even though Readers feel the opposite.

Technical Summary

1. Problem Statement

The rapid integration of Artificial Intelligence (AI) into writing processes has created a "creatorship gap," where AI-generated or AI-assisted text is often indistinguishable from human writing. This ambiguity challenges readers' ability to contextualize content, assess trustworthiness, and judge accountability. While existing regulations often mandate disclosure, there is a misalignment between top-down policies and the "grassroots" perceptions of the actual stakeholders: readers (content consumers) and writers (content creators).

The core research problem is to understand the specific factors that influence the perceived necessity of AI disclosure. Specifically, the authors investigate how the perspective (reader vs. writer), the purpose of the text (evaluation, learning, entertainment), and procedural factors (how AI was used) shape these perceptions.

2. Methodology

The study employs a vignette experiment (a survey-based experimental design using hypothetical scenarios) to systematically manipulate variables and measure perceptions.

• Participants: $N = 727$ participants recruited from Prolific (US-based, English-proficient, AI-literate).
• Design: A Resolution IV fractional factorial design was used to manage the combinatorial complexity of the variables.
• Independent Variables:
  1. Perspective: Two levels (Reader vs. Writer).
  2. Purpose: Three levels (Evaluation, Learning, Entertainment), each operationalized with two specific scenarios (e.g., job application vs. cover letter for Evaluation).
  3. Procedural Factors: Four dimensions, each with High/Low levels:
     • Replaceability: Can the writer produce similar work without AI?
     • Effortfulness: How much time/effort did the writer invest?
     • Intentionality: Did the writer steer the AI (High) or let AI lead (Low)?
     • Directness: Was AI output directly incorporated (High) or used only as reference (Low)?
• Procedure: Participants were assigned a single vignette describing a writing scenario with specific combinations of the above factors. They were asked a binary question: "Is AI disclosure necessary in this situation?" (Yes/No). They also provided open-ended justifications.
• Analysis:
  • Quantitative: Logistic regression modeling to predict the binary outcome (Necessity: Yes/No), including main effects, interaction terms (Perspective $\times$ Procedural Factors), and covariates (writing self-efficacy, AI literacy, internal judgments).
  • Qualitative: Reflexive thematic analysis of open-ended responses to supplement quantitative findings.

3. Key Contributions

1. Bottom-Up Perspective: Shifts the focus from regulatory mandates to the "perceived necessity" held by the direct stakeholders (readers and writers), addressing a gap in FAccT transparency research.
2. Synthesis of Procedural Factors: Consolidates prior frameworks (e.g., CCC framework, CoCo matrix) into four distinct, operationalizable procedural factors (Replaceability, Effortfulness, Intentionality, Directness) specific to AI-assisted writing.
3. Identification of Perspective Gap: Empirically demonstrates a significant divergence between how readers and writers perceive the need for disclosure, particularly regarding the role of intentionality.
4. Design Guidelines: Proposes concrete design interventions for AI disclosure tools, including "scaffolding" (automated, granular disclosure templates) and behavioral nudges based on specific usage patterns.

4. Key Results

The logistic regression and qualitative analysis yielded the following findings:

• Perspective Effect (RQ1):

  • Readers perceive disclosure as significantly more necessary than writers (Writers have ~80% lower odds of perceiving disclosure as necessary).
  • Interpretation: Writers may suffer from self-serving bias or view AI as an "ambient" infrastructure, reducing their sense of obligation to disclose.

• Purpose Effect (RQ2):

  • The purpose of the text (Evaluation, Learning, Entertainment) had no statistically significant main effect on the perceived necessity of disclosure.
  • Note: While qualitative data suggested some context-specific nuances, the quantitative model showed that how AI was used mattered more than why the text was written.

• Procedural Factors (RQ3):

  • Replaceability: Disclosure is seen as more necessary when AI contribution is irreplaceable (Low Replaceability).
  • Effortfulness: Surprisingly, the amount of effort put in by the writer had no significant effect on perceived necessity. This contradicts the "effort heuristic" (the idea that less effort implies less ownership).
  • Intentionality: Disclosure is seen as more necessary when the writer does not intentionally steer the AI (Low Intentionality).
  • Directness: Disclosure is seen as more necessary when AI content is directly incorporated with minimal changes (High Directness).

• Interaction Effects (RQ4):

  • Intentionality $\times$ Perspective: This interaction was significant and counterintuitive.
    • Readers: Perceive disclosure as more necessary when Intentionality is Low (AI leads).
    • Writers: Perceive disclosure as less necessary when Intentionality is Low.
  • This suggests writers may feel less ownership (and thus less need to disclose) when they let AI take the lead, whereas readers feel this lack of human control makes disclosure critical for trust.

• Covariates:

  • Higher Writing Self-Efficacy and AI Literacy correlated with a lower perceived necessity of disclosure.
  • Stronger Internal Judgment of Disclosure (valuing transparency) correlated with a higher perceived necessity.

5. Significance and Implications

The study provides a data-driven foundation for designing better AI disclosure mechanisms:

• Beyond Binary Labels: Current "AI-generated" tags are insufficient. Disclosure should be granular, detailing the replaceability, intentionality, and directness of AI use, as these are the factors readers actually care about.
• Scaffolding Disclosure: Platforms should offer tools that automatically generate disclosure drafts based on interaction logs (e.g., "You used AI to generate 80% of the text with minimal editing"). Crucially, these tools should de-emphasize effort (as it didn't matter to participants) and highlight intentionality (as this is a major point of divergence between readers and writers).
• Behavioral Nudges: To align writer behavior with reader expectations, systems should deploy targeted nudges when Low Intentionality is detected (i.e., when the writer lets AI take the lead), reminding them that this specific scenario is highly valued for transparency by readers.
• Policy Reform: Regulations should move away from rigid "all-or-nothing" rules toward context-aware guidelines that reflect the nuanced procedural factors identified in the study.

In conclusion, the paper argues that effective AI transparency requires moving beyond simple detection or mandatory labeling to a system that understands the nature of human-AI collaboration and bridges the perceptual gap between those who create the content and those who consume it.